Simulation-driven probe design and dual-wavelength optimization for wearable cerebral oxygenation monitoring

Abstract

We present a simulation-driven framework that jointly optimizes probe geometry, source–detector separation, and dual-wavelength configuration for wearable continuous-wave functional near-infrared spectroscopy. Five candidate layouts built around a non-collinear 2 × 3 matrix probe were evaluated on a Monte Carlo eXtreme platform using cortical-shell contribution ratio and detector-fluence metrics, followed by Cramér–Rao lower bound (CRLB) analysis of all ten wavelength pairs drawn from {660, 760, 810, 850, 904} nm, with tissue absorption including an explicit water term and pairs ranked directly by total estimation uncertainty (σtotal) rather than a weighted composite score. The two-dimensional cross-coverage geometry with the 660/904 nm pair yielded the lowest total estimation uncertainty (σtotal = 0.0490 μM, cond(A) = 1.50) and remained stable under pathlength, noise-model, and fluence-cutoff perturbations. A geometry-defined motor-cortex region of interest on a public finger-tapping dataset provided pipeline validation, a Colin27 anatomical-atlas simulation supported the directional robustness of the geometry result, and a four-channel prototype demonstrated hardware feasibility. The principal contribution is a reproducible simulation–statistics–validation methodology; the 660/904 nm configuration is the framework’s model-based prediction under the present tissue model, to be confirmed by future in-vivo wavelength comparison.

1. Introduction

Near-infrared spectroscopy (NIRS) and its functional counterpart (fNIRS) exploit the differential absorption of oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) within the 650–950 nm optical window to noninvasively monitor cerebral hemodynamics [1–4]. The technique offers bedside portability, sub-second temporal resolution, and tolerance to moderate subject motion, positioning it as a compelling modality for both clinical cerebral-oxygen surveillance and task-evoked functional neuroimaging.

Despite these strengths, three fundamental obstacles impede broader clinical translation of continuous-wave (CW) fNIRS. First, the modified Beer–Lambert law (MBLL) underpinning most CW concentration estimates is acutely sensitive to effective-pathlength modeling. Second, source–detector separation (SDS) governs a Pareto trade-off between cortical sensitivity and photon-flux detectability: lengthening SDS enhances depth penetration but attenuates detected flux and amplifies noise, with diminishing returns beyond a geometry-dependent threshold [5–8]. Third, dual-wavelength selection remains largely empirical in deployed systems, decoupled from any formal criterion of inversion precision; prior work has shown that the optimal pair depends jointly on site-specific tissue optical properties, attenuation, and measurement-matrix conditioning rather than on raw penetration depth [9,10]. Probe geometry and optode layout likewise continue to rely on experience rather than a unified, reproducible optimization framework [8,11].

In this work we propose a simulation-driven framework that jointly optimizes probe geometry, SDS, and dual-wavelength configuration for wearable cerebral oxygenation monitoring. Starting from a full scan over five candidate wavelengths {660, 760, 810, 850, 904} nm, we evaluate candidate layouts using the cortical-shell contribution ratio (Rcortex-shell), a detector-fluence proxy (logDetP), and a composite MBLL–Cramér–Rao lower bound (CRLB) metric. The pipeline is then externally validated on a public finger-tapping dataset [12] and deployed on a four-channel prototype. The principal novelty lies in elevating dual-wavelength selection from heuristic practice to a statistically grounded experimental-design problem, while layered external evidence closes the gap between in silico optimization and real-world feasibility.

2. Probe design and simulation methods

2.1. Non-collinear compact two-dimensional probe

Five candidate layouts were constructed to span the depth-sensitivity–detectability trade-off space (summarized in Table 4): a short-distance reference channel (1D_SD10), a one-dimensional far-distance structure (1D_Far30), a one-dimensional differential geometry (1D_Diff), and two two-dimensional layouts (2D_Horizontal and 2D_Cross). The 2D non-collinear design introduces angular diversity without substantially enlarging the scalp-contact footprint, improving cortical coverage uniformity while preserving extensibility to higher-density arrays. The probe node set is N = {a, b, c} ∪ {e, f, g}, where b and f house dual-wavelength sources and a, c, e, g are detectors; the mirror-symmetric rows yield cross-coverage on the scalp surface (Fig. 1). The four-channel prototype is a proof-of-concept realization aimed at verifying that the optimized parameters translate to a real acquisition chain, not at replicating a full high-density array.

Table 4. Primary metrics for the five candidate geometries at 660 and 904 nm (mean ± SD, N = 5).

Geometry	Rc @660 (%)	logDetP @660	Rc @904 (%)	logDetP @904	Assessment
1D_SD10	2.992 ± 0.001	5.338 ± 0.002	3.203 ± 0.000	5.364 ± 0.001	Shallow ref.
1D_Far30	13.654 ± 0.002	3.115 ± 0.005	14.350 ± 0.010	3.159 ± 0.010	Viable
1D_Diff	16.808 ± 0.007	−30.000	17.594 ± 0.008	−30.000	Undetectable
2D_Horiz.	14.377 ± 0.005	3.521 ± 0.009	15.132 ± 0.004	3.557 ± 0.009	Good balance
2D_Cross	14.695 ± 0.007	3.401 ± 0.005	15.475 ± 0.005	3.443 ± 0.006	Best Pareto balance

Fig. 1.

Schematic of the non-collinear 2 × 3 compact probe array and channel-pairing conventions.

2.2. Layered head model and tissue optical parameters

A four-layer homogeneous slab model—scalp (4 mm), skull (4 mm), cerebrospinal fluid (CSF, 2 mm), and brain (remainder)—described the photon-transport medium. Probe nodes were placed at the topmost voxel center (z = 1), consistent with skin-contact wearable mounting. Tissue optical parameters at each candidate wavelength followed Strangman et al. [8] and Jacques [13] for scalp, skull, and brain, and Fang and Boas [14] for CSF; a water-absorption term (Hale & Querry [15]) was added to each layer using literature water fractions, which is negligible at 660 nm but significant at 904 nm (Table 1). The joint optimization covered five wavelengths {660, 760, 810, 850, 904} nm; the main text presents 660/904 nm results in detail.

Table 1. Tissue optical parameters of the four-layer slab head model at 660 and 904 nm (all coefficients in mm⁻¹).

Layer	d (mm)	μa@660	μs@660	μs'@660	μa@904	μs@904	μs'@904	g	n
Scalp	4	0.017	7.4	0.74	0.023	6.4	0.64	0.9	1.4
Skull	4	0.011	9.4	0.94	0.015	8.4	0.84	0.9	1.4
CSF	2	0.004	0.3	0.03	0.011	0.3	0.03	0.9	1.4
Brain	Rest	0.018	8.4	0.84	0.022	6.7	0.67	0.9	1.4

Absorption coefficients include a water-absorption contribution μa,total = μa,tissue + f_water·μa,water(λ), with pure-water coefficients from Hale & Querry [15] and tissue water fractions (scalp 0.65, skull 0.30, CSF 0.99, brain 0.80). The water term is negligible at 660 nm but non-negligible at 904 nm (shoulder of the 970 nm water band); the largest changes are CSF (0.004→0.011) and brain (0.017→0.022) at 904 nm.

2.3. Monte Carlo simulation and Jacobian sensitivity field

Photon transport was modeled with Monte Carlo eXtreme (MCX) [14]. For every geometry–wavelength combination, five independent random-seed replicates (Nphot = 1.6 × 10⁸) were executed on the unified head model to yield mean ± SD estimates. A supplementary photon-budget convergence check confirmed that Rcortex-shell had plateaued at this setting. The Jacobian sensitivity field for each source–detector pair was constructed via the adjoint method:

$$\mathrm{J}(\mathrm{r})=\mathrm{\Phi }\mathit{\text{src}}(\mathrm{r})\cdot \mathrm{\Phi }\mathit{\text{det}}(\mathrm{r})$$ (1)

where Φsrc(r) and Φdet(r) denote the forward and adjoint fluence distributions, respectively. Depth-integrated Jacobians yielded cumulative sensitivity profiles and the cortical-shell contribution metric (Fig. 2). The cortical shell was defined as the brain region spanning z = 10–13 mm (4 mm scalp + 4 mm skull + 2 mm CSF onset, 3 mm shell thickness), approximating superficial cortical gray matter.

Fig. 2.

Four-layer head model with cortical-shell definition and cumulative depth-sensitivity profile.

2.4. Multi-objective evaluation and CRLB-based wavelength optimization

Three metric classes were employed: (i) Rcortex-shell, the fractional Jacobian contribution from the cortical shell, quantifying relative cortical sensitivity; (ii) detector-end fluence proxies logDetP and logJtot, gauging engineering signal strength; and (iii) the CRLB of joint HbO/HbR estimation derived from the MBLL measurement matrix. Let ΔOD denote the optical-density changes at two wavelengths:

$$\mathrm{\Delta }\text{OD}=\mathrm{A}\cdot \mathrm{\Delta }\mathrm{c}$$ (2)

where Δc = [ΔHbO, ΔHbR]^T and A embeds extinction coefficients and effective pathlengths. A shot-noise covariance Σ was constructed from the MCX detector fluence (reference count Nref = 10⁷), yielding the Fisher information matrix F = A^TΣ⁻¹A and the CRLB:

$$\text{Cov}(\mathrm{\Delta }\hat{\mathrm{c}})={({\mathrm{A}}^{\mathrm{T}}{\mathrm{\Sigma }}^{-1}\mathrm{A})}^{-1}$$ (3)

Individual uncertainties σHbO and σHbR were combined into σtotal, which served as the primary ranking criterion. We initially considered a weighted composite score S = 2.0·sCRLB + 1.0·sRc + 0.5·scond, but a weight-sensitivity analysis (Section 4.2) showed that this composite ranking is unstable: across a 0–3 grid of the three weights, three different pairs each become the top-ranked configuration in different regions of weight space. Because any weighted composite introduces weight-dependent arbitrariness, we rank pairs directly by σtotal, which corresponds directly to inversion precision and is independent of subjective weighting; Rcortex-shell and cond(A) are retained as supplementary descriptors. Ranking robustness was assessed under ±10% pathlength perturbation, alternative noise models, and fluence-cutoff threshold variation. Extinction coefficients used throughout are listed in Table 2.

Table 2. Extinction coefficients of HbO and HbR at each candidate wavelength (source: Cope/Matcher/Prahl).

λ (nm)	εHbO (mm−1·mM−1)	εHbR (mm−1·mM−1)
660	0.3196	3.2266
760	1.3720	1.8300
810	2.0640	1.7140
850	2.5330	1.5400
904	3.2130	0.9680

3. Data processing and validation

3.1. Public finger-tapping dataset

The publicly available Robert Luke finger-tapping BIDS/SNIRF dataset [12,16,17] (760/850 nm) served as an independent calibration source. Four processing steps were applied: (i) geometric reconstruction from SNIRF coordinates with rejection of channels whose SDS was <5 mm or deviated from nominal by >50%; (ii) conversion of raw intensity I(t) to optical-density changes relative to baseline I0,

$$\mathrm{\Delta }\text{OD}(\mathrm{t})=-{\text{log}}_{10}\left[\mathrm{I}(\mathrm{t})/{\mathrm{I}}_0\right]$$ (4)

followed by joint dual-wavelength MBLL inversion into per-channel HbO/HbR time series; (iii) bandpass filtering (0.01–0.5 Hz) and stimulus-locked block averaging to extract the hemodynamic response function (HRF); and (iv) anatomically defined ROI aggregation. To avoid selecting channels on the basis of the observed activation (double-dipping), the ROI was defined purely from channel geometry: long-distance channels (SDS ≥ 25 mm) over the left motor cortex (C3 region), selected by source–detector midpoint position (x_mid < −25 mm, |y_mid| < 35 mm, x_mid > −80 mm) without reference to any hemodynamic data. Five channels (S1-D2, S1-D3, S3-D2, S3-D3, S4-D3) met this anatomical criterion, so the block-level statistics in Section 4.3 are confirmatory rather than exploratory.

3.2. Four-channel prototype and self-collected data

The prototype employed a standard block-design finger-tapping paradigm (35 s baseline, five blocks of 25 s rest + 20 s tapping at 2 Hz visual cue, 15 s post-rest; total ≈ 225 s; Fig. 3). A dual-wavelength, four-channel synchronous receiver with HHLL time-division framing provided ambient-light suppression through differential subtraction:

$$\mathit{\text{Ieff}}=H-L$$ (5)

Fig. 3.

Temporal paradigm of the self-collected finger-tapping experiment.

Per-channel ΔOD and MBLL-derived HbO/HbR time series were processed identically to Section 3.1. Because the prototype remains at a proof-of-concept stage, results are treated as a feasibility demonstration rather than clinical-grade verification.

3.3. Statistical analysis

Following established fNIRS statisti cal-analysis practice [18], methods f or each comparison scenario are su mmarized in Table 3.

Table 3. Statistical methods for each comparison scenario.

Scenario	Test	Tail	Note
2D_Cross vs. 2D_Horiz. (Rc)	Welch t + Mann–Whitney U	One-sided	N = 5 seeds
Public-data ROI block effect	One-sample t + Wilcoxon	One-sided	n = 30 blocks; geometry-defined ROI
Prototype data	Descriptive only	—	Proof-of-concept

4. Results

4.1. Multi-objective comparison of candidate geometries

Within the two 2D geometries, 2D_Cross showed a marginally higher Rcortex-shell than 2D_Horizontal at both wavelengths: the 660 nm difference was 0.318 percentage points (Welch t, p = 2.47 × 10⁻¹²; Mann–Whitney, p = 0.004); the 904 nm difference was 0.343 pp (p = 1.53 × 10⁻¹⁴; Mann–Whitney, p = 0.004). Although the Welch t-test returns extremely small p-values, this reflects only the determinism of Monte Carlo replicates sharing identical anatomy under very high photon counts (SD ≈ 0.007%) and does not represent a physiologically meaningful advantage; we therefore regard the ∼0.3 pp difference as practically negligible and its real-world import as requiring multi-subject evaluation. On the strength of its favorable Pareto position (Fig. 4), 2D_Cross was nonetheless carried forward as the reference geometry for wavelength optimization, with the anatomical robustness of this choice examined on the Colin27 atlas (Section 5).

Fig. 4.

Pareto front of Rcortex-shell versus logDetP for candidate probe-geometry variants at 660 nm.

4.2. Dual-wavelength optimization

All ten pairwise wavelength combinations were evaluated under the 2D_Cross geometry (Table 5, Figs. 5, 6).

Table 5. Top five dual-wavelength pairs ranked by σtotal (2D_Cross, SDS = 33 mm).

Rank	λ pair (nm)	Rc (%)	cond(A)	σtotal (μM)
1	660 / 904	12.02	1.50	0.0490
2	660 / 850	12.15	2.04	0.0567
3	660 / 810	12.20	2.50	0.0678
4	760 / 904	11.98	3.39	0.0951
5	660 / 760	12.20	3.86	0.1000

Fig. 5.

Heatmap of Rcortex-shell across five wavelengths and candidate probe strategies.

Fig. 6.

Pareto front of σtotal versus Rcortex-shell for all ten wavelength pairs.

The 660/904 nm pair owes its CRLB-based ranking to three physical factors: strongly complementary HbO/HbR extinction coefficients (εHbO = 0.320 vs. 3.213 mm−1·mM−1; εHbR = 3.227 vs. 0.968 mm−1·mM−1) that minimize chromophore cross-talk; the lowest measurement-matrix condition number among all pairs (cond(A) = 1.50), ensuring stable inversion; and preservation of the relative sensitivity gain of 904 nm at cortical depth alongside a stable contrast baseline at 660 nm. Crucially, this CRLB-based ranking advantage stems from the well-conditioned extinction matrix—a spectral property independent of the head model and of water absorption—so 660/904 retains the lowest σtotal (0.0490 μM) even after the water correction. Robustness analysis confirmed that the 660/904 nm ranking persisted under ±10% effective-pathlength perturbation, alternative noise models, and fluence-cutoff switching (Kendall τ = 1.000 across all scenarios). A composite four-dimensional performance heatmap is provided in Supplement 1 ^{(3.4MB, pdf)} , Fig. S1. The full ten-pair wavelength ranking is provided in Supplement 1 ^{(3.4MB, pdf)} , Table S2.

To justify the use of σtotal as the sole ranking criterion, we performed a weight-sensitivity analysis of the composite score. Across a 0–3 grid of the three weights, the composite winner was 660/810 in 53.5% of combinations, 660/904 in 24.9%, and 660/850 in 21.6%—three different pairs each rank first depending on the weighting, and under the water-corrected properties the nominal weights happen to favor 660/850 on the composite score. In contrast, the σtotal ranking is invariant to weighting: 660/904 ranks first in all cases and remains rank 1 under every perturbation (Kendall τ = 1.000). We therefore base the recommendation on σtotal alone; the composite score is retained only as an exploratory descriptor whose weight-sensitivity is documented in Supplement 1 ^{(3.4MB, pdf)} .

4.3. External validation on public finger-tapping data

The public dataset employs 760/850 nm (σtotal ≈ 0.142 μM under the present framework); this section therefore validates the processing pipeline, not the wavelength conclusion. After geometric correction and short-distance removal, 20 long-distance channels remained (mean SDS = 38.23 ± 2.18 mm (Fig. 8)). Five channels meeting the anatomical (geometry-defined) C3 ROI criterion (mean SDS = 39.50 ± 1.76 mm) were aggregated, independently of the observed activation.

Fig. 8.

Statistical summary of the public dataset: channel-level ΔHbO/ΔHbR, SDS distribution, and HRF peak characteristics.

The geometry-defined ROI recovered a significant task-related HbR decrease (Fig. 7), consistent with left motor-cortex activation during right-hand tapping (HbR trough −0.168 μM at 11.77 s; Wilcoxon p < 0.001). HbO did not reach significance, consistent with the common observation that HbR provides a more robust activation marker than HbO in finger-tapping fNIRS, HbO being more affected by global systemic physiology [19]. Block-level statistics are summarized in Table 6.

Fig. 7.

ROI-averaged HRF (red: HbO; blue: HbR; shading: ± SEM, n = 30 blocks).

Table 6. Block-level task-minus-baseline statistics on the geometry-defined C3 ROI (n = 30 blocks).

Metric	Effect ± SD (μM)	SEM	t-test p	Wilcoxon p
ΔHbO	+0.0215 ± 0.1464	0.0267	0.4286	0.7499
ΔHbR	−0.1006 ± 0.1859	0.0339	0.0060	0.0008

4.4. Prototype system verification

The 2D_Cross / 660–904 nm configuration was deployed on the four-channel prototype under the standard finger-tapping paradigm. Block-averaged traces (Fig. 9) exhibited stimulus-locked hemodynamic trends consistent with the canonical HbO increase / HbR decrease pattern, providing preliminary evidence that the simulation-optimized parameters operate in real hardware. Continuous multichannel time series are provided in Supplement 1 ^{(3.4MB, pdf)} , Fig. S2. Additional same-subject resting-state recordings under 660/904 and 760/850 nm are provided in Supplement 1 ^{(3.4MB, pdf)} , Fig. S8, documenting stable in-vivo optical signals and identifiable physiological pulsation under both wavelength configurations. The prototype task recordings and same-subject resting-state recordings are both limited to feasibility demonstration; they should not be interpreted as clinical validation or as definitive wavelength-comparison evidence.

Fig. 9.

Block-averaged HRF from the prototype finger-tapping experiment (mean across 5 blocks; channel S1-λD1 to S2-λD2).

5. Discussion

Among the five candidate geometries, 2D_Cross occupies the most favorable Pareto position rather than dominating on any single metric: the short-distance reference offers maximal flux but negligible cortical sensitivity; the differential structure maximizes Rcortex-shell but suffers catastrophic fluence underflow; one-dimensional far-distance layouts gain cortical contribution at rapidly escalating flux penalties. The 2D_Cross geometry leverages angular diversity to improve cortical spatial sampling while avoiding the severe attenuation of extreme separations. On wavelength selection, the 660/904 nm pair ranked first in the CRLB analysis because its extinction-coefficient matrix is well-conditioned and its CRLB is minimized through complementary noise weighting—counter to the naïve expectation that deeper-penetrating wavelengths are inherently superior. The layered validation strategy—public data for pipeline calibration, prototype for hardware translation—mitigates the closed-loop bias of self-validated systems and forms a simulation–calibration–demonstration evidence chain.

The reproducibility of the proposed framework should therefore be interpreted at the workflow level rather than as universal generalization across all instruments or populations. The reproducible components are the geometry construction, MCX simulation workflow, Jacobian extraction, CRLB-based wavelength ranking, sensitivity analysis, and geometry-defined public-data ROI processing. The present evidence does not yet generalize to all anatomies, devices, subjects, hardware noise conditions, probe-contact states, or clinical populations. Generalization requires matched optical properties, valid source-detector geometry, adequate detected fluence, consistent extinction coefficients, and a verified acquisition chain under the intended measurement conditions.

To test whether the geometry result depends on the simplified slab model, we re-simulated the 2D_Cross and 2D_Horizontal geometries on the Colin27 anatomical atlas (181 × 217 × 181, 1 mm isotropic, seven tissue types) at 660 nm, with optodes placed on the C3 scalp surface (Fig. 10) and cortical sensitivity defined anatomically as the Jacobian fraction within gray matter rather than a fixed-depth shell. The 2D_Cross advantage was preserved: Rcortex was 4.20 ± 0.001% for 2D_Cross versus 3.64 ± 0.001% for 2D_Horizontal (+0.57 pp), the same direction as in the slab model (14.70% vs 14.38%, +0.32 pp). Because the two models use different cortex definitions, the absolute Rcortex values are not directly comparable; only the sign of the Cross−Horizontal difference is compared, and both models agree that the angular-diversity advantage of the cross geometry survives realistic cranial curvature and CSF light-piping.

Fig. 10.

Colin27 optode placement for the 2D_Cross geometry at C3 (660 nm validation), shown on three orthogonal atlas slices; all six optodes lie on the scalp surface.

Five limitations warrant acknowledgment. First, the homogeneous four-layer slab model neglects inter-individual anatomical variability; subject-specific three-dimensional models [20] and alternative layered models [21] would strengthen the generalizability of the findings. Second, the CRLB analysis assumes MBLL linearity and shot-noise dominance, which may under-represent complex in vivo noise sources; future work should incorporate physiological noise and motion-artifact models [22]. Third, in the public-data validation, the ROI was defined purely from channel geometry over the motor-cortex region, independently of the observed activation, thereby avoiding the double-dipping concern associated with activation-based ROI selection. Fourth, the prototype data are limited to a single subject and remain susceptible to attachment-state instability; systematic multi-subject and multi-session validation is therefore required before clinical translation. Fifth, and most importantly, although an additional same-subject resting-state measurement was obtained after replacing the source configuration with the conventional 760/850 nm pair, this exploratory recording does not constitute a definitive wavelength-comparison experiment. The two configurations were not acquired under a fully matched multi-session task paradigm with identical raw-data formats, repeated probe attachment, and multi-subject replication. Therefore, the present 660/904 nm result is still framed as the model-based prediction of the proposed CRLB framework rather than as an experimentally established in vivo optimum. A rigorously matched multi-subject, multi-session comparison of 660/904 and 760/850 nm on identical hardware remains the primary priority for future validation.

6. Conclusion

This work presented a non-collinear 2 × 3 compact probe and a closed-loop evaluation framework for wearable fNIRS cerebral oxygenation monitoring. Monte Carlo simulations identified 2D_Cross as the geometry with the most favorable Pareto balance (Rcortex-shell = 15.475% at 904 nm), while CRLB analysis identified 660/904 nm as the pair with the lowest total estimation uncertainty (cond(A) = 1.50, σtotal = 0.0490 μM, rank 1 under all perturbations, Kendall τ = 1.000), a ranking driven by spectral conditioning and therefore robust to the water-absorption correction. A geometry-defined motor-cortex ROI on a public finger-tapping dataset recovered a significant task-related HbR response, a Colin27 anatomical simulation supported the geometry ranking direction, and a four-channel prototype demonstrated hardware feasibility. The primary contribution is not an isolated parameter set but a reproducible simulation–statistics–validation methodology for probe and wavelength optimization; the 660/904 nm configuration is the framework’s prediction under the present tissue model, awaiting confirmation by future multi-subject in-vivo wavelength comparison.

Funding

Hebei Provincial Department of Education https://ror.org/01jkyjd96 ( CXZZDD202602); S&T Program of Hebei ( 26240801D).

Disclosures

The authors declare no conflicts of interest.

Data availability

The public finger-tapping dataset underlying Section 4.3 is available via the Robert Luke fNIRS BIDS/SNIRF repository [12,17]. MCX simulation code and analysis scripts used in this study are available from the corresponding author upon reasonable request.

Supplemental document

See Supplement 1 ^{(3.4MB, pdf)} for supporting content.